allus-company-data (Python)

The Python SDK for the allus company-data API. Point it at a JSON config file and it hands back typed, plaintext, your-slug-keyed conclusions: for each connected person, a map of your request-field slug → plaintext value (plus whether the value is live and when it last changed).

The SDK hides everything else — the OAuth token, the field catalog, the id plumbing, the hybrid decryption, binary fetching, the changes-queue mechanics, JSON-vs-XML. The platform is zero-knowledge: the API only ever holds ciphertext, so all decryption happens inside the SDK with your service private key. The person's own field choices are never exposed — you only ever see the request slots you configured.

This SDK is one of six language ports that share an identical API surface. This manual is the Python view of it.

Contents: TL;DR — fetch new updates · Quickstart · Every call · The typed value model · The changes pump · Webhooks · Company documents · Contract-flow runs · Rate limits · Errors · How it's wired

Deeper reference pages live in docs/: config · model · pump · webhooks · errors.

TL;DR — fetch new updates

A system/Homebrew Python refuses a bare pip install (PEP 668) — install into a virtualenv:

python3 -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install allus-company-data

Point a config.json at your service keys:

{
  "api_url": "https://api.allme.fyi",
  "client_id": "svc_xxx",
  "client_secret": "xxx",
  "service_private_key": "/path/to/service.pem",
  "key_passphrase": "xxx",
  "cache_dir": "./allus-cache"
}

Drain everything new, handled one update at a time:

from allus_company_data import Client

client = Client.from_config("config.json")

def handle(change):
    # one update at a time: event, person, slug, value, live, at
    print(change.event, change.person_id, change.slug, change.value,
          "live" if change.live else "snapshot", change.at)

client.process_changes(handle)   # returns when the feed is empty

process_changes pulls every pending change, decrypts it, and hands them to your callback ONE BY ONE, acking each only after your code returns. Crash mid-batch? The next run replays exactly what wasn't acked — nothing is lost, and the API keeps no backlog of its own. Run it on a schedule (cron / systemd timer); there is no daemon/follow mode by design. Connections, binary values, and webhooks are documented below.

Quickstart

Requires Python ≥ 3.11. A system/Homebrew Python refuses a bare pip install (PEP 668) — install into a virtualenv:

python3 -m venv .venv && source .venv/bin/activate   # Windows: .venv\Scripts\activate
pip install allus-company-data
# or, working from a clone:    pip install -e '.[dev]'      # from the repo root
python -c "import allus_company_data; print(allus_company_data.__version__)"

1. Write a config file

A single JSON file holds everything. Any field can be overridden by an ALLUS_* env var, so secrets needn't live in the file. No SDK method ever takes a key, passphrase, or secret as an argument — they all come from here.

allus.json:

{
  "api_url": "https://api.allme.fyi",
  "client_id": "svc_1a2b3c…",
  "client_secret": "…",
  "service_private_key": "./service-CRM.pem",
  "key_passphrase": "…",

  "account_private_key": "./account.pem",
  "account_passphrase": "…",

  "webhooks": {
    "wh_abc123": "hmac_secret_for_that_webhook"
  },

  "cache_dir": "./allus-cache",
  "format": "json"
}

Env overrides use the ALLUS_ prefix of the field name, e.g. ALLUS_CLIENT_SECRET, ALLUS_KEY_PASSPHRASE, ALLUS_ACCOUNT_PASSPHRASE, ALLUS_WEBHOOK_SECRET. A missing/invalid config (or an unreadable PEM / wrong passphrase) raises ConfigError at construction — fail fast.

2. First call — list a connection's values

from allus_company_data import Client

client = Client.from_config("allus.json")

# Iterate every connected person (lazy, auto-paged).
for conn in client.connections():
    print(conn.display_name, conn.person_id)
    for slug, val in conn.values.items():
        print(f"  {slug} = {val.value!r}  (live={val.live}, updated={val.updated_at})")
    break  # just the first one for the demo

Or fetch one connection by id:

conn = client.connection("019xxxxxxxxxxxxxxxxxxxxxxxxx")
email = conn.values["work_email"].value        # "alice@acme.com"  (a str)

client = Client.from_env() builds the same client entirely from ALLUS_* env vars (no file).

Every call

Client is the only object you construct. Build it from config, then:

Client.from_config(path, **kwargs) -> Client     # from a JSON file (env overrides secrets)
Client.from_env(**kwargs)          -> Client      # entirely from ALLUS_* env vars

kwargs are advanced/optional: http (an injected HttpClient), logger (a logging.Logger), sleep (a Callable[[float], None], for tests).

request_fields()

request_fields() -> list[RequestField]

Your request-field definitions — fetched once from GET /api/company-data/request-fields and cached for the life of the client (it types every value). Returns your request config, never the person's fields.

• Params: none.
• Returns: list[RequestField] — each RequestField(slug, label, type, one_time, mandatory, raw). mandatory is true when the field is mandatory-to-provide or mandatory-to-stay-connected.
• Raises: AuthError, ApiError, RateLimitError.

for f in client.request_fields():
    flag = "mandatory" if f.mandatory else "optional"
    print(f"{f.slug:20} {f.type:10} {flag}{' (one-time)' if f.one_time else ''}")

connections(limit, offset)

connections(limit: int = 100, offset: int = 0) -> Iterator[Connection]

A lazy generator that auto-pages GET /api/company-data/connections?limit&offset and yields one typed Connection at a time (bounded memory for a large book). Each conn.values[slug] is already decrypted (or a lazy binary handle).

• Params: limit — page size (default 100); offset — starting offset.
• Returns: Iterator[Connection].
• Raises: AuthError, ApiError, DecryptError (per value, at access), RateLimitError (after the iterator's bounded internal backoff — see Rate limits).

Heavily rate-limited. Use for the initial full sync + occasional reconciliation only — never as a poll substitute for the changes feed. The generator paces itself within the limit (backs off on Retry-After).

# Initial full sync, streaming so a 100k-connection book never lands in memory.
for conn in client.connections(limit=200):
    upsert_local_record(conn)

connection(id)

connection(id: str) -> Connection

Fetch one connection by its connection id (GET /api/company-data/connections/{id}).

• Params: id — the connection id (Connection.id).
• Returns: one Connection. Note: this endpoint returns {connection_id, user_id, values} and no display_name/connected_at, so those identity fields are None here (the list endpoint carries them).
• Raises: AuthError, ApiError (404 if unknown), DecryptError, RateLimitError.

conn = client.connection(conn_id)
phone = conn.values.get("mobile")
if phone:
    print(phone.value, "live" if phone.live else "snapshot")

logs(limit, offset)

logs(limit: int = 50, offset: int = 0) -> list[LogEntry]

The service's activity log (GET /api/company-data/logs?limit&offset) — ops events only (email / purge / webhook), never person field data.

• Params: limit (default 50), offset (default 0).
• Returns: list[LogEntry] — each LogEntry(type, message, metadata, at, raw).
• Raises: AuthError, ApiError, RateLimitError.

for entry in client.logs(limit=20):
    print(entry.at, entry.type, entry.message)

process_changes(handler, **options)

process_changes(handler: Callable[[Change], None], **options) -> None

The crash-safe changes pump: drains the feed through handler one Change at a time, durably buffering each batch before delivery, with per-item ack and retry → dead-letter → continue. Runs until the feed is empty, then returns — there is no follow/daemon mode (you schedule re-runs yourself). Delivery is at-least-once, so your handler must be idempotent (dedup on Change.id). See The changes pump for the full model.

• Params: handler — your callback; called with one Change. A return is an ack; an exception triggers retry.
• Options (keyword-only): batch_size (clamped to ≤ 500, default 100), max_retries (default 3), on_error ("deadletter" — default — or "halt"), backoff (Callable[[int], float], attempt → seconds).
• Returns: None (when the feed is empty + the buffer is drained).
• Raises: AuthError, ApiError, RateLimitError (during a drain); ValueError (bad on_error); whatever the handler raises if on_error="halt" and retries are exhausted.

def handle(change):
    if already_processed(change.id):      # idempotency — dedup on the stable id
        return
    if change.event == "field_updated":
        store(change.person_id, change.slug, change.value)
    elif change.event in ("connection_deleted", "field_deleted"):
        remove(change.person_id, change.slug)
    mark_processed(change.id)

client.process_changes(handle)            # returns when the feed is empty

logger is not a process_changes option in this SDK — pass it once to the Client constructor (Client.from_config("allus.json", logger=my_logger)).

Advanced changes primitives

drain_batch(max: int = 100)                      -> list[Change]   # raw, UNBUFFERED — you own durability
dead_letters()                                   -> list[dict]      # the local dead-letter store
retry_dead_letters(handler, **options)           -> int             # re-drive dead-lettered events; returns count re-driven

• drain_batch(max) — fetches one batch (clamped ≤ 500) and returns the decrypted Changes directly. It does not persist anything, so a crash loses what the API already deleted. Prefer process_changes for safe consumption.
• dead_letters() — each dict is the stored (ciphertext) event plus a flattened error and attempts.
• retry_dead_letters(handler, **options) — same max_retries / on_error / backoff options as process_changes; on success a record is removed, on repeated failure it stays dead-lettered (or re-raises under "halt"). Dead letters are never re-fetched from the API — the local store is their only home.

for dl in client.dead_letters():
    print("stuck:", dl["id"], dl["error"], "after", dl["attempts"], "attempts")

n = client.retry_dead_letters(handle)     # after you've fixed the bug
print(f"re-drove {n} dead letters")

Key rotation — key_rotated and the public-key cache

Every client caches the RSA public keys it fetches: a person's key is immutable — until they rotate it. A person learns of a rotation from a silent push; your service gets no pushes, so the key_rotated change is your only signal. Without it a long-running worker keeps encrypting to the rotated-away key for its whole lifetime, and the person can never read those values.

On the pump this is automatic — the cached key is dropped as the change passes through, before your handler sees it. Over a webhook it is not: the signature verifier is static and has no client instance, so it cannot reach the cache. Call the invalidator yourself — noting that the two clients key their caches differently: the service client by share_code, the customer client by the person's user id. Passing a share code to the customer client removes nothing and leaves you encrypting to the old key. Both identifiers ride every change, alongside public_key_sha256 — the fingerprint of the person's new key.

if change.event == "key_rotated":
    client.invalidate_public_key(change.share_code)      # service Client — keyed by SHARE CODE
    customer.invalidate_public_key(change.person_id)     # CustomerClient — keyed by PERSON USER ID
    # change.public_key_sha256 = fingerprint of the NEW key, if you want to verify the refetch

This is eventual, not fail-closed — nothing rejects a document encrypted to a stale key, so a window remains between the rotation and your next drain. Drain often if that window matters.

service_key_rotated — the same thing, the other way round

The customer client also caches the service's public key, the one you encrypt your consent answers and documents to, keyed "companyCode/serviceCode". When that company replaces its service keypair, the service_key_rotated change on your account feed is your only signal — you receive no pushes. Same shape, same guarantees, same automatic handling on the pump:

if change.event == "service_key_rotated":
    # Automatic on the pump. Over a webhook, from the raw event body:
    customer.invalidate_service_key(body["company_share_code"], body["service_share_code"])
    # body["service_public_key_sha256"] = fingerprint of the service's NEW key

Also eventual, not fail-closed. Note the identifiers are share codes, not the ids used by invalidate_public_key — the two caches are keyed differently and the wrong call removes nothing.

Webhook helpers (on the client)

The webhook receiver helpers are also exposed as Client methods (they delegate to the module functions, fully config-driven — no key/secret arguments):

client.verify_webhook(raw_body: bytes, headers: dict) -> bool
client.parse_webhook(raw_body: bytes, headers: dict)  -> Change
client.handle_webhook(raw_body: bytes, headers: dict) -> Change   # verify + parse

• verify_webhook — recomputes HMAC-SHA256(raw_body, secret) and constant-time-compares it to X-Allus-Signature. Returns True/False; never raises for a bad signature.
• parse_webhook — body → a typed Change. Does not verify. Handles JSON, XML, and the encrypt_payload account-key envelope. Raises WebhookError on a malformed/unparseable body.
• handle_webhook — verify then parse; raises WebhookError on a bad/unknown signature, otherwise returns the Change. The typical one-liner inside a route.

The same three are importable as standalone functions (from allus_company_data import verify_webhook, parse_webhook, handle_webhook), which take the config and the decrypt/type closures explicitly — but inside an app you'll almost always use the client methods. See Webhooks.

The typed value model

You work with these objects and nothing else (from allus_company_data import …):

RequestField { slug, label, type, one_time, mandatory, verified, verified_max_age_days }
Connection   { id, person_id, display_name, connected_at, values: {<slug>: Value} }
Value        { value, live, updated_at, verified, verified_at, verified_expires_at }
Change       { id, event, person_id, slug?, value?, live?, document_id?, status?, at }
Document     { id, kind, name, description, status, payload_kind, is_private, value, metadata, created_at, updated_at }
LogEntry     { type, message, metadata, at }

Keyed by your slug

conn.values["work_email"].value → "alice@acme.com". The key is the stable, explicit slug you set per request field in the portal — rename the label freely, the slug is the contract. The person's source field is never exposed: no source slug, no field_id, not even via .raw.

Value(value, live, updated_at)

Value types (from the field's type)

country/nationality values are 2-letter ISO codes, and an address's country/state sub-fields are an ISO alpha-2 code / USPS 2-letter state code respectively. is_field_value_valid(type, value) validates these against the bundled country dataset; is_valid_country_code(code) / dial_code_for(code) check a code or look up its E.164 dial code.

addr = conn.values["home_address"].value     # dict, e.g. {"street": "...", "city": "...", ...}
dob  = conn.values["birthday"].value          # datetime.date(1990, 5, 17)

Binary fields — the lazy BinaryHandle

A photo/document value is a BinaryHandle. Nothing is fetched or decrypted until you call .bytes() or .save():

handle = conn.values["passport_scan"].value   # BinaryHandle (no network yet)

data = handle.bytes()                          # GET the slot file → the file bytes
n    = handle.save("/tmp/passport.jpg")        # same, written to disk; returns bytes written
print(handle.value_url)                         # the opaque slot-keyed URL it fetches from
print(handle.content_type)                      # what the bytes arrived as, once fetched
print(handle.content_sha256)                    # the platform's digest of exactly those bytes

.bytes() GETs the slot-keyed file endpoint and returns the file bytes — but that endpoint has two 200 shapes, and which one you get is the person's choice, not yours. It depends on whether their source field is private, they can change that at any time, and nothing announces it in advance:

• private source → application/json, {"encrypted": true, "value": <wrapper>}. The handle decrypts the wrapper with your service key, parses the inner JSON envelope ({"full": "data:…"} for photos, {"file": "data:…"} for documents) and base64-decodes the data URI into the file bytes.
• plaintext source → the file's own Content-Type (image/jpeg, application/pdf, …) and the body IS the file. Nothing is decrypted and no service key is needed.

The handle hides the difference: .bytes()/.save() give you the file either way. The shapes are told apart on the response Content-Type, never by looking at the body — a PDF that happened to start with a brace must not be mistaken for a wrapper. The result is cached on the handle, so repeated calls don't re-fetch.

Every 200 carries X-Allus-Content-Sha256, the sha256 of exactly the bytes returned; handle.content_sha256 is that header (and handle.content_type the Content-Type), so you can record what you received and later show your archived copy has not drifted. It is the platform's word, not a signature. There is no variant selection — one slot has one byte sequence and therefore one digest.

A frozen (share-once) answer is retained for 90 days. After that the endpoint returns 410 company_data.file_expired, which surfaces as an ApiError whose details carry the answer's content_sha256 and expired_at — your archived copy is then the only one, and you can still prove what it is:

try:
    data = handle.bytes()
except ApiError as e:
    if e.error_key == "company_data.file_expired":
        log(e.details["content_sha256"], e.details["expired_at"])

Change(id, event, person_id, slug?, value?, live?, at)

A change-feed / webhook event.

.raw

Every model carries .raw — the underlying hardened API dict — for debugging or an edge case the SDK didn't model. It still never contains the person's source field.

See docs/model.md for the full reference.

The changes pump

The changes feed is a server-side drain-on-fetch queue: GET /api/company-data/changes?limit=N returns up to N events (default 100, max 500) and deletes exactly those rows in the same transaction — no offset/cursor, and the API keeps no copy afterward. So consumption can't be a plain list: a consumer crash mid-batch would lose events the API already deleted, and a huge backlog must not materialize in memory. process_changes solves both.

Per run, repeating until the feed is empty then returning:

1. Replay first. Deliver any un-acked events already in the local buffer (from a previous crashed run), oldest-first.
2. Drain. When the buffer is empty, fetch one batch and persist it to the durable file buffer (fsync) BEFORE handing anything out. This is the backup the API no longer has.
3. Deliver one-by-one. For each buffered event, oldest-first: decrypt its value at delivery (never on disk), build the typed Change, call handler.
4. Ack / retry / dead-letter. On success, remove the event from the buffer (ack). On a handler error, retry with backoff up to max_retries; then either move it to the dead-letter store and continue (on_error="deadletter", default — one poison event never wedges the stream) or stop and re-raise (on_error="halt"). A DecryptError on a buffered event (corrupt/truncated ciphertext, rotated key) is dead-lettered immediately — re-decrypting can't fix it, so it does not burn retries (under on_error="halt" it re-raises). Either way it never propagates out and wedges replay.
5. Repeat until a drain returns empty and the buffer is drained → return.

The durable buffer

• Plain files under cache_dir (zero extra dependencies): pending/ for un-acked events, deadletter/ for ones that exhausted retries.
• Stored events keep their ciphertext value — no plaintext PII is ever written to disk. Decryption happens only at delivery.
• Writes are crash-safe (temp file → fsync → atomic rename → dir fsync). Files are named with a monotonic, zero-padded sequence so they replay oldest-first.

Crash safety, at-least-once, and idempotency

A batch is durably buffered before any delivery, and acked per-item only after the handler succeeds. The ack can't be atomic with your side-effects — a crash between your handler's success and its ack re-delivers that event on the next run. That makes delivery at-least-once, so:

Your handler must be idempotent. Dedup on Change.id.

Change.id is the stable server change-row id, captured before the server delete, so it survives crash + replay unchanged.

No follow mode

process_changes returns when the feed empties. You schedule re-runs — a cron job, a while True: client.process_changes(handle); time.sleep(5) loop, a worker queue, whatever fits. The feed is cheap to poll (see Rate limits).

Worked example

import time
from allus_company_data import Client

client = Client.from_config("allus.json")

def handle(change):
    # Idempotent: skip anything we've already applied.
    if seen(change.id):
        return
    match change.event:
        case "field_updated":
            store_value(change.person_id, change.slug, change.value, live=change.live)
        case "field_deleted":
            clear_value(change.person_id, change.slug)
        case "connection_deleted":
            drop_person(change.person_id)
        case "connection_created" | "consent_accepted" | "consent_declined":
            note_event(change.person_id, change.event, change.at)
    record_seen(change.id)

# Schedule your own re-runs; process_changes itself returns when empty.
while True:
    client.process_changes(handle, batch_size=200, max_retries=5)
    time.sleep(5)

If a handler keeps failing, the event lands in the dead-letter store instead of blocking the stream; inspect with client.dead_letters() and re-drive with client.retry_dead_letters(handle) after fixing the cause. See docs/pump.md.

Webhooks

Webhooks are the lower-latency push alternative to polling the changes feed. The platform POSTs each change event to your configured webhook URL with:

• X-Allus-Webhook-Id — which webhook this is (selects the HMAC secret from config).
• X-Allus-Signature — HMAC-SHA256(rawBody, secret) as lowercase hex.
• the body — the same slug-keyed Change shape as the pull feed (JSON or XML).

All secrets/keys come from config; the helpers take no key or secret arguments. Use the raw request body bytes (do not re-serialize a parsed body — the HMAC is over the exact bytes the platform sent).

Delivery contract — effectively unique, rarely replayed

Each queued event is POSTed once, and only HTTP 200 counts as delivered — a 202, a 204, a 3xx redirect and every 4xx/5xx are all treated as a failure. On anything other than 200 (or a timeout or connection error) the event is not retried in place: it and the rest of the webhook's queue move to a durable server-side backlog and the webhook is marked bad. The backlog is delivered later, either automatically when the webhook next probes healthy, or when you drain it yourself with GET /api/company-data/changes?webhook_id=… (delete-on-read).

So deliveries are effectively unique — with one rare exception. If your endpoint processed an event but the platform never saw your 200 (your response timed out, or you crashed after committing but before responding), the event is treated as failed and replayed on recovery, so you receive it again. Nothing caps that at two: a failed probe leaves its backlog row in place, so every later recovery attempt whose 200 is likewise lost replays the same event once more. Inside that window the contract is at-least-once — plan for one or more repeats, not for exactly one.

Do not use change.id as an idempotency key here. On the webhook path the id is neither reliably stable nor reliably fresh, and a receiver cannot tell which one it is holding. A live delivery is built with no change row behind it, so its id is minted for that single POST — the later replay of the same event is rebuilt from a durable backlog row and therefore carries a different id. But a replayed delivery carries that row's id, and the row stays in place until it is delivered successfully, so a re-attempted replay arrives with the same id — which changes again if the event is re-backlogged after a further failure. An id check therefore misses the duplicate you are most likely to see and matches only a rarer one; it is not a contract. If you need strict idempotency, key on the content — event + person + slug/document + payload — never on the id.

Webhooks and the pull feed are alternative integrations — consume one, never both. The id-dedup guidance in the changes-pump section above applies to the pump only, where change.id is the real server change-row id.

In a web route (Flask)

from flask import Flask, request, abort
from allus_company_data import Client, WebhookError

app = Flask(__name__)
client = Client.from_config("allus.json")

@app.post("/allus/webhook")
def allus_webhook():
    try:
        change = client.handle_webhook(request.get_data(), dict(request.headers))
    except WebhookError:
        abort(401)              # bad / unknown signature, or unparseable envelope

    # Do NOT carry the pump's id-dedup over here: the webhook id is not an
    # idempotency key (see "Delivery contract" above). Key on content if you need one.
    apply_change(change)
    return ("", 200)   # 200 — the ONLY status allus counts as delivered

verify_webhook / parse_webhook let you split the steps if you prefer:

if not client.verify_webhook(raw_body, headers):
    abort(401)
change = client.parse_webhook(raw_body, headers)

Config-driven secrets

Per-webhook HMAC secrets live in the config webhooks map, keyed by webhook id; the SDK reads X-Allus-Webhook-Id off the request and looks up the matching secret. A single-webhook service can use the flat "webhook_secret": "…" shortcut (or ALLUS_WEBHOOK_SECRET). An unknown/unconfigured id ⇒ verification returns False (and handle_webhook raises WebhookError).

The encrypt_payload account-key envelope

If a webhook has encrypt_payload enabled, the body is replaced by a {"_enc":1,…} envelope encrypted to your company account key (and the HMAC is over that envelope — the final bytes sent). parse_webhook/handle_webhook unwrap it transparently using the configured account_private_key + account_passphrase, then decrypt the inner field value with the service key — so an encrypted-payload Change is identical to a plain one. If you receive such a webhook without an account_private_key configured, you get a WebhookError.

The account-key envelope uses OAEP-SHA1 (OpenSSL's default), distinct from the OAEP-SHA256 used for person field values — the SDK handles this difference internally; you only supply the account key in config.

See docs/webhooks.md.

Company documents

Documents are content your service issues to people (a quote, a contract, a JSON payload, a PDF) — the mirror image of the request slots. They come in two shapes:

• Broadcast — no target. Sent to every connection on the service. Plaintext — you can't single-key-encrypt one value to all your connections, so a broadcast value is stored as-is.
• Per-person — targeted at one connection (connection_id= / person_user_id= / share_code=). Automatically end-to-end encrypted to that recipient's public key before it leaves the process. The server only ever stores ciphertext.

The encryption rule, plainly: every per-person document is automatically end-to-end encrypted to the recipient's public key — for any is_private value. Broadcast documents are plaintext. is_private is a device-display-only flag (it picks lock-screen vs decrypt-on-load on the recipient's device), not what decides encryption — so is_private=True with no per-person target raises ConfigError. As everywhere in this SDK, no method ever takes a key or secret argument — the recipient key is fetched for you, and your own service key comes from config.

Creating documents

payload_kind picks the body:

• payload_kind="json" — pass json_value= (a JSON-serializable object).
• payload_kind="file" — pass file_bytes= (+ file_mime=); the bytes are uploaded. For a per-person file the bytes are encrypted automatically too.

from allus_company_data import Client

client = Client.from_config("allus.json")

# 1. BROADCAST plaintext json doc — every connection sees it, no target.
notice = client.create_document(
    kind="document",
    name="2026 price list",
    payload_kind="json",
    json_value={"plan": "pro", "monthly": 49, "currency": "EUR"},
)

# 2. PER-PERSON doc — auto-encrypted to that recipient's public key.
#    Target it by ANY one of connection_id / person_user_id / share_code.
contract = client.create_document(
    kind="document",
    name="Service agreement",
    payload_kind="json",
    is_private=True,                       # display-only; needs a per-person target
    connection_id="019xxxxxxxxxxxxxxxxxxxxxxxxx",
    json_value={"tier": "enterprise", "term_months": 12},
)

# A per-person FILE — the bytes are encrypted for the recipient automatically.
with open("agreement.pdf", "rb") as fh:
    pdf = client.create_document(
        kind="legal_document",
        name="Signed agreement",
        payload_kind="file",
        person_user_id="019yyyyyyyyyyyyyyyyyyyyyyyyy",
        file_bytes=fh.read(),
        file_mime="application/pdf",
    )

# is_private without a target → ConfigError (a broadcast can't be locked).

Listing, reading, updating, deleting

list_documents(*, person_user_id=None, status=None, limit=100, offset=0) -> list[Document]
document(document_id)                                                     -> Document
document_file(document_id)                                                -> bytes    # #491: the file BYTES
flow_run_document(run_id)                                                 -> bytes    # #491: the run's OWN (service-key) copy
update_document_status(document_id, status)                              -> Document
update_document_metadata(document_id, *, metadata=None, name=None,
                         description=None)                               -> Document
delete_document(document_id)                                            -> None

# All this service's documents (optionally filter by person and/or status).
for doc in client.list_documents(status="offering"):
    print(doc.id, doc.name, doc.status, doc.payload_kind, "private" if doc.is_private else "shared")

doc = client.document(contract.id)

# For a json doc, .json() returns the plaintext object — a per-person doc is
# decrypted with your service key on demand; a broadcast doc is already plaintext.
print(doc.json())

# #491: download a FILE document's bytes (metadata methods don't include them).
# A BROADCAST document is served plaintext and is returned as-is. A PER-PERSON /
# private document is encrypted to the RECIPIENT's key — the company cannot
# decrypt that, so this raises ApiError('documents.recipient_encrypted') instead
# of a doomed decrypt attempt.
try:
    pdf_bytes = client.document_file(contract.id)
except ApiError as e:
    if e.error_key == "documents.recipient_encrypted":
        # This is a per-person document; only the recipient can read it. For a
        # generated flow contract, download the company's OWN copy instead:
        pdf_bytes = client.flow_run_document(run.id)
    else:
        raise

# Move it through its lifecycle / edit its metadata.
client.update_document_status(contract.id, "active")
client.update_document_metadata(contract.id, name="Service agreement (v2)",
                                metadata={"renewal": "auto"})

client.delete_document(notice.id)

A Document carries id, kind, name, description, status, payload_kind, is_private, value, metadata, created_at, updated_at (and .raw). Use .json() on a payload_kind="json" document to get the decrypted plaintext object.

Reacting to a status change in the feed

When someone advances one of your documents (e.g. signs it), the platform emits a document_status_changed change. In a process_changes handler it carries .document_id and .status (and no slug/value — it's a lifecycle event, not a field value):

def handle(change):
    if change.event == "document_status_changed":
        # the document moved lifecycle state (offering → ready_to_sign → active → …)
        on_document_status(change.document_id, change.status)
    elif change.event == "field_updated":
        store(change.person_id, change.slug, change.value)

client.process_changes(handle)

Messaging

Your service can hold a conversation with a connected person — the same messaging surface the person already uses, with your service as the counterpart. Two shapes:

• 1-on-1 — send_message(connection_id, text). End-to-end encrypted: the SDK encrypts one copy to the person's public key and one to your service key before anything leaves the process, so the person reads it in their app and you can re-read your own outbound text. The platform stores ciphertext only.
• Broadcast — broadcast_message(text). One plaintext message to every person connected to the service (one body cannot be single-key-encrypted to all of them, exactly as for a broadcast document). It seeds each person's ordinary 1-on-1 thread; their reply comes back end-to-end encrypted.

send_message answers 201 with the created message carrying message_id, and returns that id — the value you hand back as the acknowledgement boundary.

Inbound messages arrive on the changes pump / webhook as a message_received event — a person→company message only. A broadcast raises no event of its own.

def handle(change):
    if change.event != "message_received":
        return
    print(change.person_id, change.message_body)   # already decrypted for you

    # Reply on the same connection. person_public_key rides the event, so no
    # second key lookup is needed.
    client.send_message(
        change.connection_id,
        "Thanks — we're on it.",
        person_public_key=change.person_public_key,
    )

    # Acknowledge what you handled. REQUIRED: without it the message stays
    # unread forever, your unread count grows, and the person never sees a read
    # receipt. Sending a reply does NOT acknowledge anything.
    client.mark_messages_read(change.connection_id, change.message_id)

client.process_changes(handle)

mark_messages_read is bounded by the boundary message: a message that arrived while you were working is not swept, and a repeat is a no-op. The boundary must be a message the person sent on that connection — anything else is refused with ApiError("company_data.ack_boundary_invalid") (400).

# One plaintext announcement to everyone connected to the service.
client.broadcast_message("We're closed on Friday.")

Refusals surface as ApiError carrying the platform error_key:

Contract-flow runs (company side)

The company is one bound party of a contract flow — a multi-step, per-party form the platform walks to gather (and end-to-end encrypt) answers, optionally finishing at a document-generating leaf. These calls cover the company's turn:

trigger_flow_run(flow_id, *, connection_id, bindings)      -> FlowRun
flow_runs(*, status="awaiting_company")                    -> list[FlowRun]
flow_run(run_id)                                            -> FlowRun
flow_run_answers(run)                                        -> dict[str, str]  # #491 gap 1
submit_flow_answers(run, fill, *, party_pubkeys=None)       -> FlowRun
generate_flow_document(run)                                  -> dict
process_flow_run(run_id, fill_node, *, party_pubkeys=None)  -> FlowRun
identity()                                                    -> dict           # #491 gap 3

• trigger_flow_run(flow_id, connection_id=..., bindings={...}) starts a run bound to a connection and the flow's other parties, pinning the flow's latest published version.
• flow_runs(status=...) / flow_run(run_id) list / fetch runs. status=None returns everything; the default "awaiting_company" is the actionable queue.
• flow_run_answers(run) (#491 gap 1) — a run's decrypted answers as {slug: plaintext}, reading the company's service-key answer copies. Accepts a loaded FlowRun or a run id (fetched via flow_run).
• submit_flow_answers / generate_flow_document / process_flow_run fill the company's current node, advance the run (encrypting one answer copy per bound party), and — at a document-mode leaf — generate the contract. See the method docstrings for the full per-party encryption details.
• identity() (#491 gap 3) — this client's own {"company_user_id": ..., "service_id": ...} from GET /api/company-data/whoami. trigger_flow_run's company-side binding must use company_user_id (the person party's user_id comes from the connection) — without this call it was unconstructible through the SDK.

me = client.identity()
run = client.trigger_flow_run(
    flow_id, connection_id=conn.id,
    bindings={"company": me["company_user_id"], "person": conn.person_id},
)

# Later, once the run is complete:
answers = client.flow_run_answers(run.id)     # {slug: plaintext} — the company's copies

# If the flow's output_mode is "document", download the company's OWN generated
# copy (encrypted to the SERVICE key, unlike a per-person document_file()):
pdf_bytes = client.flow_run_document(run.id)  # see Company documents above

• Raises: AuthError, ApiError (404 on flow_run/flow_run_document for an unknown run, or one with no generated document yet), DecryptError, RateLimitError, ValidationError (from submit_flow_answers on a slug failing field-type validation).

Rate limits

A 429 carries Retry-After. The SDK backs off and retries automatically:

• The transport (HttpClient) retries a 429 a bounded number of times honoring Retry-After, then surfaces RateLimitError.
• The connections(...) generator additionally backs off per Retry-After on a surfaced RateLimitError and retries the page a bounded number of times before re-raising — so it paces itself within the limit instead of hammering.

If you catch a RateLimitError, its .retry_after is the seconds to wait (or None when the header was absent).

Your client_credentials token requests (/oauth2/token) are on their own rate-limit bucket, separate from person logins — but it is keyed by source IP, not by your client_id, so it is shared with every other client_credentials caller reaching the API from the same address (another service on your network, a second client on the same host). Caching the token, as described under How it's wired below, is what keeps that shared window from being spent needlessly — by you or anyone else behind the same IP. Every rate-limit refusal — this 429, and the platform's 503 when its own limiter store is unreadable — now carries a populated .error_key, readable off the same RateLimitError/ApiError.

Errors

All from allus_company_data. Same taxonomy + names across all six SDKs.

from allus_company_data import (
    Client, ConfigError, AuthError, ApiError,
    DecryptError, WebhookError, RateLimitError,
)

try:
    client = Client.from_config("allus.json")
    for conn in client.connections():
        ...
except ConfigError as e:
    ...   # fix the config / key file
except RateLimitError as e:
    wait(e.retry_after or 60)
except ApiError as e:
    log(e.status, e.error_key, e.message)

See docs/errors.md.

How it's wired

Everything below is what the SDK hides so your code only ever sees conclusions.

Auth / token. An HttpClient owns a client_credentials-only token. On the first call (or when the cached token nears expiry) it POSTs client_id/client_secret to {api_url}/oauth2/token and caches the bearer token + its expiry; refresh is automatic. A mid-flight 401 triggers exactly one refresh-and-retry, then AuthError. The token is scoped server-side to one service, so every call is implicitly that service's data.

Regions. The configured api_url is the platform's global front door and also the starting point for every request, including the token request. A client_credentials token is minted at your company's home region, and the token response names that region's base in an api_url member — the SDK stores it and sends every subsequent request there, token requests included, because the token is valid only at that region and a company that moves region is followed by the next mint. A data call that still reaches the front door is refused with 421 + error_key: region.rebase_required, carrying the same api_url; the SDK stores it and retries the call exactly once. The SDK does not validate a server-returned api_url against anything — it stores the base the server names and uses it. An absent or empty api_url is never stored and the response surfaces as the error it is.

Slug resolution. request_fields() is fetched once and cached; its slug→type map types every value (so address parses to a dict, photo becomes a lazy binary handle, etc.). The connection/changes endpoints return values keyed by your request slug — the person's source field is dropped server-side and never reaches the SDK.

Decryption (zero-knowledge). The service private key is loaded once at construction from the configured encrypted PEM + passphrase into an in-memory RSA key. A decrypt closure over it is handed to every model factory and the pump — the key never appears in a method signature. Each value is a hybrid wrapper ({"_enc":1,"k":rsa_oaep_sha256(aesKey),"iv":…,"d":aes256gcm(…)}); the SDK RSA-OAEP-SHA256 unwraps the AES key, then AES-256-GCM decrypts the payload. The platform only ever holds ciphertext — it never sees your plaintext.

Binary fetch. A binary value is a lazy BinaryHandle over a slot-keyed value_url. On .bytes()/.save() it GETs that file endpoint and classifies the response on its Content-Type: an encrypted answer ({"encrypted":true,"value":…}) runs the same service-key decrypt to a JSON file-envelope and base64-decodes its data URI, while a plaintext answer's body already IS the file. Either way you get the file bytes, plus the response's X-Allus-Content-Sha256 on .content_sha256. (Slot-keyed, never source-field-keyed.)

The drain-on-fetch feed. process_changes delegates to a Pump wired to a fetch_changes closure (GET /changes?limit=, returning raw ciphertext events) and a decrypt closure (builds a typed Change). Because the fetch deletes the rows it returns, the pump persists each batch to the durable file buffer (ciphertext at rest) before delivery, acks per-item after your handler succeeds, and replays the buffer on restart — see The changes pump.

Sign in with allme (OAuth, #195)

Relying-party helper for the "Sign in with allme" identity flow. Config-only keys (the idw role):

from allus_company_data import OAuthClient, Claim

oauth = OAuthClient.from_config("idw-config.json")  # {api_url, oauth_client_id, oauth_redirect_uri, oauth_client_secret?, oauth_private_key?, oauth_key_passphrase?}
url = oauth.authorize_url("signin", state="xyz", code_challenge=challenge)   # the button target
# ...user approves; your redirect_uri receives ?code=...
info = oauth.complete_sign_in(code, code_verifier=verifier)  # {user, mode, values(plaintext), values_cipher}

Modes: signin (identity), one_time (frozen claim values, decrypted for you), connect (a lasting connection), 2fa_enroll (opt a person into 2FA — see below). authorize_url(mode, claims=[Claim("email", "email", suggest="email_personal")]); poll_result(state) for the detached response mode.

#498 — a claim IS a request field. You describe what you need and the person picks which of their own fields answers it; you never name a field. A claim carries a mandatory unique name (everything that comes back is keyed by it — values, values_cipher, attestations, and their stored choice for a repeat login), a field type, an optional suggested slug, required, and verified ("only a #311-verified answer will do"). A nameless or duplicate claim raises ConfigError at the call rather than failing at the API. verified is accepted only on the OIDC flow and only for a type allme can verify (today email); elsewhere it is refused with invalid_request rather than quietly dropped.

verified_max_age_days narrows a verified claim to a RECENT verification, and the merge is tighten-only: the app's registered configuration is a FLOOR, a request may only tighten it, and the effective limit is the minimum of the two stated ages. An omitted age tightens nothing — omitting it sends nothing at all, never an explicit null — and a value below 1 raises ConfigError at the call.

complete_sign_in returns {user, mode, two_factor, values, values_cipher, attestations}.

• user.sub is the person's share code and equals share_code — byte-identical to the id_token's sub. display_name is gone: ask for a name claim and read values["name"].
• values_cipher is an additive sibling of values, keyed the same way: the raw app-key ciphertext wrapper each plaintext value was decrypted from, exactly as userinfo delivered it. Lets you show that a value really came from encrypted delivery rather than trusting it verbatim. Empty for a mode/claim that carries no ciphertext (signin, or plaintext delivery) — that emptiness is the honest answer.
• attestations is an additive sibling map keyed by the same claim name, present only for a verified claim under encrypted delivery. Each entry carries a verified boolean the SDK computes itself, in constant time, over the plaintext it just decrypted — plus the raw hash/salt/verified_at/ verified_expires_at. A slug ABSENT from the map is "not attested", never "wrong" (treat that value as unverified); an entry present with verified false is a MISMATCH and you must reject the value. verified_at attests the value as verified at that moment, not verified today; verified_expires_at is when that verification lapses on its own (None = it does not), and an expired attestation is unverified — the computed verified already reads false once it has passed.

resolve_userinfo(access_token, fallback_mode=None) is the second half of complete_sign_in — the userinfo read + decrypt + attest, without the token exchange — for a caller whose exchange already ran through a different client (a standards-only third-party OIDC library, say, that verified the id_token itself but cannot read a claim value the id_token never carries). Config-only key handling applies exactly as it does to complete_sign_in: you pass no key or passphrase, only the access token you already hold. Returns the identical shape (values, values_cipher, attestations) and carries the same mismatch-rejection duty on the caller. complete_sign_in is implemented on top of this method. fallback_mode is used only when userinfo itself omits mode — pass the mode your own token response carried, or None if you have none.

2FA by allme (#436, #481)

Ask a connected person to approve a login inside the allme app. On the same service data client (no new config), via the two_factor sub-client:

from allus_company_data import Client

client = Client.from_config("allus.json")

# Raise a challenge. idempotency_key is REQUIRED — a repeat within the TTL returns the SAME challenge and
# sends no second push. `context` is plain text shown on the person's card.
ch = client.two_factor.challenge("2I6UF3", idempotency_key="login-8f3c1a", context="Sign-in from Chrome")
if ch.matching_digits:                        # number matching is on for this service
    show_on_login_page(ch.matching_digits)    # the person types these back into the app; the server checks them

# Wait for the terminal outcome — polls result() for you, raises ApiError on timeout.
res = client.two_factor.wait_for_result(ch.challenge_id)   # or result(ch.challenge_id) to poll once yourself
if res.status == "approved":
    grant_login()

• Burn-on-read. The first read of a terminal state (approved | denied | expired | revoked) delivers it and burns it — a later read is gone. Read it once and persist your own outcome; wait_for_result returns that first terminal read and never re-reads a consumed challenge.
• Webhook variant. The 2fa_challenge_completed change/webhook carries the same terminal status, so a webhook consumer need not poll. Expiry fires no webhook/Change — only approved/denied/revoked reach the feed, so a lapsed challenge is observable only by polling.
• Enrollment. Only an enrolled person can be challenged (an un-enrolled share_code is 404). Enrollment is a one-time consent on the web.allme.fyi/auth surface via the OAuth helper's 2fa_enroll mode — a redirect button (oauth.authorize_url("2fa_enroll", state=...)), or server-to-server with response_mode="detached" + poll_result(state), which returns {"enrolled": true, "state": ...} once the person confirms.
• Errors. 404 (unknown / not-enrolled share code). A 429 is either the plain rate limit (retried with backoff → RateLimitError) or twofa.pending_cap (too many challenges already open for this person) — the latter surfaces immediately as ApiError and is never retried, since a retry cannot clear it.